youtube-shorts-generator

YouTube Shorts Generator

End-to-end pipeline that turns one long video into N viral-ready vertical clips. Each clip ships with a viral score (0–100), an opening hook line, and a one-sentence reason it should perform.

Reference implementation: https://github.com/SamurAIGPT/AI-Youtube-Shorts-Generator

When to use this skill

• "Generate shorts from this YouTube video"
• "Find the most viral 60-second clips in this podcast"
• "Auto-crop this interview to 9:16"
• "Give me TikTok clips from this lecture"

If the user only wants transcription, summarization, or thumbnails — this is the wrong skill.

Inputs to collect before running

Ask once, then proceed:

1. Source — YouTube URL (preferred) or path/URL to an mp4
2. num_clips — default 3
3. aspect_ratio — default 9:16 (also: 1:1, 4:5)
4. language — default auto-detect (forwarded to MuAPI Whisper as ISO-639-1)
5. Output JSON path — optional; if set, dump full result there

If the user gave a URL and nothing else, use defaults and don't block on questions.

Prerequisites (verify before first run)

• Python 3.10+
• A MuAPI key — set MUAPI_API_KEY in .env. Powers download, transcription, highlight ranking, and clipping. If missing, stop and ask the user for it; do not invent one.
• pip install -r requirements.txt inside a venv

If the repo isn't cloned yet, clone https://github.com/SamurAIGPT/AI-Youtube-Shorts-Generator.git into the working directory.

Pipeline (what to execute)

Run the eight stages in order. Each maps to a module in shorts_generator/.

1. Download (downloader.py) — pull the source video at the requested resolution (360/480/720/1080, default 720).
2. Transcribe (transcriber.py) — MuAPI /openai-whisper runs Whisper server-side and returns timestamped verbose_json segments. Billed per minute of audio.
3. Classify content type — LLM tags the video (podcast / interview / tutorial / vlog / lecture / monologue) and density. Tune the highlight prompt per type.
4. Chunk if long (highlights.py) — videos > LONG_VIDEO_THRESHOLD (1800s default) are split into CHUNK_SIZE_SECONDS (1200s default) windows with CHUNK_OVERLAP_SECONDS (60s default) overlap so cross-boundary highlights aren't missed.
5. Rank highlights — LLM scans each chunk through VIRALITY_CRITERIA:
   • Hook moments — strong opening line that stops the scroll
   • Emotional peaks — laughter, anger, vulnerability, awe
   • Opinion bombs — spicy, contrarian, debate-bait takes
   • Revelation moments — "wait, what?" reframes
   • Conflict — disagreement, tension, callouts
   • Quotable lines — tight, screenshot-worthy phrasing
   • Story peaks — climax of a narrative arc
   • Practical value — actionable insight a viewer will save Each candidate gets start_time, end_time, score 0–100, title, hook_sentence, virality_reason. Aim for 30–75s clips unless content dictates otherwise.
6. Dedupe — collapse overlaps. Rule: if two candidates overlap > 50%, keep the higher score, drop the other.
7. Top-N selection — sort surviving candidates by score, take num_clips.
8. Vertical auto-crop (clipper.py) — render each highlight at aspect_ratio. Auto-handles face tracking and screen recordings; no Haar cascades.